NO_CUDA_EXT=1

To use LMCache-Ascend on the NPU hardware

Mount-VHD

--max-ttft 10.0

--cache-hit-rates

The main thing is that the working-set-size easily fits in the “left over” capacity of HBM KV Cache capacity above.

Qwen3-Coder-30B-A3B-Instruct-FP8

20.62x

ensure all prefix matches come from the base vLLM prefix matching

ShareGPT Dataset

export HCCL_OP_EXPANSION_MODE="AIV" # Enable FlashComm_v1 optimization when tensor parallel is enabled. export VLLM_ASCEND_ENABLE_FLASHCOMM1=1

CANN Toolkit

LMCacheAscendConnector

LMCacheAscendConnectorV1Dynamic

Kimi-2.5

For instance, it occasionally retries after making an initial mistake

We attribute this failure mode to RDT-1B's "Alternating Condition Injection" scheme, which alternates between injecting visual inputs and language inputs in successive transformer layers — a design that encourages the policy to pay more attention to language inputs rather than over-relying on visual inputs.

第 9 講：影像和聲音上的生成策略

TESTSPARK

HITS

SYMPROMPT

TESTPILOT

TELPA

MUTAP

default

ecdict-ultimate-mobi.zip

ecdict-ultimate-eudic.zip

ecdict-ultimate-eudic-headless.zip

ecdict-ultimate-sqlite.zip

ecdict-ultimate-stardict.zip

we also build ChatUniTest Mod-els[13 ], which provides fine-tuned models for Java test generationtasks based on Code Llama

set up ctags to understand Go.

RQ4: How is the correctness of the unit tests generated by TELPA?• RQ5: How generalizable is TELPA?

AI 辅助编程的所有提升都是基于给 LLM(大语言模型) 传递更合适的上下文这个基本原理展开的，无论是 Chat, RAG, Rules, MCP 还是未来更酷的一些其它技术

intel xpu

Building effective agents

In our study, developers accepted around 30% of GitHub Copilot’s suggestions

In our survey, respondents most commonly reported using the time they save with AI coding tools to design systems, collaborate, and learn.

nearly all of the survey participants reported using AI coding tools both outside of work or at work at some point

Easier to work with new programming languages, and understand existing codebases.

AutoDidact: Bootstrapping Search Through Self-Verification

Instead of continuously increasing pre-training budgets, test-time compute allows modes to “think longer” during inference

Proposal Distribution Enhancement

First, SFT’s methodology of minimizing the discrepancybetween predicted outputs and gold-standard references in-herently caps model performance at the quality level of thetraining data.

Thismethod systematically explores the repository knowledge graphand prioritizes the discovery of critical information such as reposi-tory functions and dependency structures that have a greater impacton resolving issues. By simulating multiple trajectories and evaluat-ing their importance, MCTS dynamically narrows the search spaceand focuses computational resources on the most relevant areas.

Excessive reference relationships may increase the complexity ofthe graph structure and affect the analysis efficiency and accu-racy of the model.

The main role of SBFL-identified methods is to reveal more “hints”on relevant classes and methods beyond those mentioned in theproblem statement. The LLM agent can then use the context re-trieval APIs to examine these methods. Since the SBFL-identifiedmethods are presented to the agent together with the problemstatements, the agent can then cross-reference between these twosources of information.

To this end, we develop a novel ASE method namedRepoUnderstander by guiding agents to comprehensively under-stand the whole repositories. Specifically, we first condense thecritical information of the whole repository into the repositoryknowledge graph in a top-to-down mode to decrease the complex-ity of repository. Subsequently, we empower the agents the abilityof understanding whole repository by proposing a Monte Carlotree search based repository exploration strategy. In addition, tobetter utilize the repository-level knowledge, we guide the agents tosummarize, analyze, and plan. Then, they can manipulate the toolsto dynamically acquire information and generate the patches tosolve the real-world GitHub issues.

To identify the essential code elements needed to com-plete the given infilling method m in a repository, a naivesolution might scan the entire codebase for all accessibleelements, which would introduce excessive noise. Anotherapproach could focus on methods with similar signatures orcontexts; however, these often provide irrelevant elementsthat do not serve m’s functional purpose, leading to redun-dancy and missing critical elements.

pruning the specific implementations of func-tions in all dependent files does not signifi-cantly reduce the accuracy of completions

Greedy Selection. Retrieval is performed if<cc> is the most likely token following <eof>.• Threshold Selection. If the probability of <cc>